ESDS Program

Advancing an Open-Access Repository for Earth Observation Training Data and Machine Learning Models

Principal Investigator: Kevin Booth (Radiant Earth)

Radiant Earth has developed an open-access repository for Earth observation training data and machine learning (ML) models. As part of the ACCESS-19 project, we have focused on three major improvements. First, we produced a multi-mission global land cover training dataset. Second, we started development on an open application programming interface (API) for registering and retrieving machine learning models. Lastly, we developed a Python client to reduce the complexity of retrieving machine learning training datasets and models from our open-access repository.

Project Objectives

  • Produce a multi-mission global land cover training dataset
  • Develop an Open API for registering and retrieving ML models
  • Develop a Python client to increase usability of Radiant MLHub

Update

In late 2019, Radiant Earth started development of our open-access repository for Earth observation training data and machine learning models. As part of the ACCESS-19 cooperative agreement, we developed a multi-mission global land cover training dataset and expanded the features available on Radiant MLHub. LandCoverNet, our multi-mission global land cover training dataset, comprises a total of 8,941 image chips of 256 x 256 pixels. This training dataset was generated from 300 geographically diverse tiles of Sentinel-2 imagery and covers the regions of Africa, Asia, Australia and Oceania, Europe, North America, and South America. Alongside the labeled land cover chips is a yearly time series of matching Sentinel-1, Sentinel-2, and Landsat-8 imagery. This dataset was first published in 2020 and has since become one of our most downloaded datasets.

Radiant MLHub has been expanded to include the ability to publish and retrieve machine learning models. These models have been cataloged in our STAC API and relevant model documentation, weights, and code can be accessed through our web interface. We developed a Python client that enables users to programmatically search for, access, and download machine learning training datasets.Radiant Earth has begun development on the next generation of Radiant MLHub, called Source Cooperative, which better supports extremely large training datasets and expands the dataset publishing capabilities beyond machine learning training datasets. Source Cooperative is in a private beta and is expected to be in public beta in Q3 of 2023.

Major Accomplishments

  • Produced a multi-mission global land cover training dataset
  • Expanded Radiant MLHub to include machine learning model publishing capabilities
  • Developed a Python client for interacting with the Radiant MLHub API

For more information

Radiant ML Hub. Current production site, to be deprecated in Q3 2023.

Publications and Presentations

Radiant Earth (2020). LandCoverNet Africa: A Geographically Diverse Land Cover Classification Training Dataset. doi:10.34911/rdnt.d2ce8i

Radiant Earth (2022). LandCoverNet Asia: A Geographically Diverse Land Cover Classification Training Dataset. doi:10.34911/rdnt.63fxe5

Radiant Earth (2022). LandCoverNet Australia: A Geographically Diverse Land Cover Classification Training Dataset. doi:10.34911/rdnt.0vgi25

Radiant Earth (2022). LandCoverNet Europe: A Geographically Diverse Land Cover Classification Training Dataset. doi:10.34911/rdnt.7s12zu

Radiant Earth (2022). LandCoverNet North America: A Geographically Diverse Land Cover Classification Training Dataset. doi:10.34911/rdnt.jx15e8

Radiant Earth (2022). LandCoverNet South America: A Geographically Diverse Land Cover Classification Training Dataset. doi:10.34911/rdnt.6a27yv

Last Updated